A Study on Eliminating Biased Node in Federated Learning

Federated Learning (FL) is a method to perform a deep learning technique with distributed nodes without aggregating their individually collected data. Each node can keep its own dataset secret, and a central server aggregates the updated parameters in a deep learning model from each node. However, it is reported that FL is vulnerable to poisoning attacks on training data. In this study, we propose a method to exclude biased nodes by analyzing the statistical characteristics of the weight parameters sent by each node to the central server. In the computer simulation, we evaluate the detection accuracy of biased node, and conduct discussions.